We published a paper where we captured the same kind of insights (deep numa hierarchies including cache levels, numa nodes, packages) and used them to tailor spinlocks to the underlying machine: https://dl.acm.org/doi/10.1145/3477132.3483557